AI-ready Data Management Practices:

Alaska Ice Seals and Harbor Seals

Stacie Koslovsky

2024-05-17

Polar Ecosystems Program

  • NOAA Fisheries - Alaska Fisheries Science Center

  • Harbor seals and 4 species of ice-associated seals

  • Abundance and distribution of all species

Image Data Streams

All image processing handled through Kitware’s VIAME software:

  • Ice Seal Surveys

  • Glacial Harbor Seal Surveys

  • Coastal Harbor Seal Surveys

KAMERA System

Kitware Image Acquisition ManagER and Archiver

Ice Seal Surveys

  • Use in-flight KAMERA system to collect infrared, RGB, and UV imagery.

  • Use detection model to identify hotspots in the thermal imagery.

  • Currently, manually review all thermal hotspots for species identification, but the development/evaluation of a color classification model is underway.

Glacial Harbor Seal Surveys

  • Use in-flight KAMERA system to collect infrared and RGB imagery.

  • Two types of surveys:

    • For lower-altitude surveys, use detection model to find hotspots in the thermal imagery (using thermal model developed for ice-associated seals). Review color imagery to classify harbor seals.

    • For higher-altitude surveys, use suppression and ignore zones to indicate areas of overlap. Manually review all images and annotate all seals. Duplicates are flagged during post-processing.

Coastal Harbor Seal Surveys

  • Collect oblique imagery in-flight using digital SLR camera.

  • Images are reviewed to label them as counted, uncounted, or background.

  • Once images are selected, run R code to generate image lists for selected images and manually review and annotate seals.

AI-Ready Data Efforts

  • Annotate all seals and subset as appropriate for population analyses.

  • Standardize annotation labeling for ease of cross-project training (for staff) and for combining datasets for different AI/ML needs.

  • Store standardized outputs from each project schema as queries (not all data were collected/processed the same way over the years, so this gives us equivalent data structures to work with).

  • Store image names, network paths and image-information data in DB structure (allows for ad hoc queries).

  • Track how images are/were used (manual review, training, test, validation, background (no seals)).

  • Develop functions in pepDataConnect R package for generating image lists and/or annotation files for use in DIVE.

Thank You

  • Aerial survey leads (Erin Moreland, Josh London, John Jansen) + program staff

  • Kitware (VIAME and KAMERA development)